A novel multi-task causal representation learning approach for interpretable maritime collision severity prediction

Wang, M ORCID: 0000-0002-3704-9858, Wang, Y ORCID: 0000-0001-8855-9770, Ma, Z ORCID: 0009-0002-6572-2912 and Wang, J ORCID: 0000-0003-4646-9106 (2026) A novel multi-task causal representation learning approach for interpretable maritime collision severity prediction. Reliability Engineering and System Safety, 271. ISSN 0951-8320

Preview

Text Accepted version.pdf - Accepted Version Available under License Creative Commons Attribution. Download (4MB) | Preview

Abstract

Accurate prediction and robust interpretation of maritime collision severity are crucial. Prevailing correlation-based methods are non-robust and lack interpretability, struggling with confounding factors, data heterogeneity, and class imbalance. A novel multi-task causal representation learning framework (MCLF) is proposed to address these limitations. Its core is a structured disentanglement mechanism that decomposes effects into direct effects and indirect effects mediated by unsafe factors, reinforced by adversarial training and orthogonality constraints to reduce representation-level confounding induced by observed covariates, thereby improving robustness and interpretability. To address class imbalance, an interactive data synthesis module using the tabular denoising diffusion probabilistic model (TabDDPM) is used, which generates high-quality samples for hard-to-classify cases to enhance model robustness. A dynamic multi-task fusion strategy then adaptively integrates the primary severity prediction with auxiliary tasks (pollution, property loss, and death). This holistic approach achieves superior predictive accuracy and enhances interpretability by providing a structurally-grounded decomposition of effects, advancing towards more transparent decision-support in maritime safety.

Item Type:	Article
Uncontrolled Keywords:	Collision severity; Causal representation learning; Class imbalance; Multi-task fusion; 40 Engineering; 49 Mathematical Sciences; 35 Commerce, Management, Tourism and Services; Machine Learning and Artificial Intelligence; Networking and Information Technology R&D (NITRD); 01 Mathematical Sciences; 09 Engineering; 15 Commerce, Management, Tourism and Services; Strategic, Defence & Security Studies; 35 Commerce, management, tourism and services; 40 Engineering; 49 Mathematical sciences
Subjects:	Q Science > QA Mathematics > QA76 Computer software T Technology > TA Engineering (General). Civil engineering (General)
Divisions:	Engineering
Publisher:	Elsevier
Date of acceptance:	12 January 2026
Date of first compliant Open Access:	29 April 2026
Date Deposited:	28 Apr 2026 10:16
Last Modified:	29 Apr 2026 00:50
DOI or ID number:	10.1016/j.ress.2026.112236
URI:	https://researchonline.ljmu.ac.uk/id/eprint/28462

View Item